From 55851275baff739df07df5a05ac541667d663094 Mon Sep 17 00:00:00 2001
From: Chuck Walbourn <chuckw@windows.microsoft.com>
Date: Sat, 18 Jun 2016 16:00:43 -0700
Subject: [PATCH] Split SSE3 optimizations from SSE4

---
 Inc/DirectXMath.h       | 51 +++++++++++++++++-----------------
 Inc/DirectXMathMisc.inl | 61 +++++++++++++++--------------------------
 2 files changed, 48 insertions(+), 64 deletions(-)

diff --git a/Inc/DirectXMath.h b/Inc/DirectXMath.h
index 40a0d8c..89de84a 100644
--- a/Inc/DirectXMath.h
+++ b/Inc/DirectXMath.h
@@ -51,11 +51,9 @@
 #define _XM_F16C_INTRINSICS_
 #endif
 
-#ifdef _XM_F16C_INTRINSICS_
-#ifndef _XM_AVX_INTRINSICS_
+#if defined(_XM_F16C_INTRINSICS_) && !defined(_XM_AVX_INTRINSICS_)
 #define _XM_AVX_INTRINSICS_
 #endif
-#endif // _XM_F16C_INTRINSICS_
 
 #if !defined(_XM_AVX_INTRINSICS_) && defined(__AVX__) && !defined(_XM_NO_INTRINSICS_)
 #define _XM_AVX_INTRINSICS_
@@ -65,7 +63,11 @@
 #define _XM_SSE4_INTRINSICS_
 #endif
 
-#if defined(_XM_SSE4_INTRINSICS_) && !defined(_XM_SSE_INTRINSICS_)
+#if defined(_XM_SSE4_INTRINSICS_) && !defined(_XM_SSE3_INTRINSICS_)
+#define _XM_SSE3_INTRINSICS_
+#endif
+
+#if defined(_XM_SSE3_INTRINSICS_) && !defined(_XM_SSE_INTRINSICS_)
 #define _XM_SSE_INTRINSICS_
 #endif
 
@@ -87,38 +89,37 @@
 #include <malloc.h>
 #pragma warning(pop)
 
-#if defined(_XM_SSE_INTRINSICS_)
-#ifndef _XM_NO_INTRINSICS_
-#include <xmmintrin.h>
-#include <emmintrin.h>
-#endif
-#elif defined(_XM_ARM_NEON_INTRINSICS_)
 #ifndef _XM_NO_INTRINSICS_
 #pragma warning(push)
 #pragma warning(disable : 4987)
 // C4987: Off by default noise
 #include <intrin.h>
 #pragma warning(pop)
+
+#ifdef _XM_SSE_INTRINSICS_
+#include <xmmintrin.h>
+#include <emmintrin.h>
+
+#ifdef _XM_SSE3_INTRINSICS_
+#include <pmmintrin.h>
+#endif
+
+#ifdef _XM_SSE4_INTRINSICS_
+#include <smmintrin.h>
+#endif
+
+#ifdef _XM_AVX_INTRINSICS_
+#include <immintrin.h>
+#endif
+
+#elif defined(_XM_ARM_NEON_INTRINSICS_)
 #ifdef _M_ARM64
 #include <arm64_neon.h>
 #else
 #include <arm_neon.h>
 #endif
 #endif
-#endif
-
-#if defined(_XM_SSE4_INTRINSICS_) && !defined(_XM_NO_INTRINSICS_)
-#pragma warning(push)
-#pragma warning(disable : 4987)
-// C4987: Off by default noise
-#include <intrin.h>
-#pragma warning(pop)
-#include <smmintrin.h>
-#endif
-
-#if defined(_XM_AVX_INTRINSICS_) && !defined(_XM_NO_INTRINSICS_)
-#include <immintrin.h>
-#endif
+#endif // !_XM_NO_INTRINSICS_
 
 #include <sal.h>
 #include <assert.h>
@@ -1625,7 +1626,7 @@ template<uint32_t SwizzleX, uint32_t SwizzleY, uint32_t SwizzleZ, uint32_t Swizz
 // Specialized swizzles
 template<> inline XMVECTOR      XM_CALLCONV     XMVectorSwizzle<0,1,2,3>(FXMVECTOR V) { return V; }
 
-#if defined(_XM_SSE4_INTRINSICS_) && !defined(_XM_NO_INTRINSICS_)
+#if defined(_XM_SSE3_INTRINSICS_) && !defined(_XM_NO_INTRINSICS_)
 template<> inline XMVECTOR      XM_CALLCONV     XMVectorSwizzle<0,0,2,2>(FXMVECTOR V) { return _mm_moveldup_ps(V); }
 template<> inline XMVECTOR      XM_CALLCONV     XMVectorSwizzle<1,1,3,3>(FXMVECTOR V) { return _mm_movehdup_ps(V); }
 #endif
diff --git a/Inc/DirectXMathMisc.inl b/Inc/DirectXMathMisc.inl
index 653dfd9..ab93620 100644
--- a/Inc/DirectXMathMisc.inl
+++ b/Inc/DirectXMathMisc.inl
@@ -1981,54 +1981,37 @@ inline XMVECTOR XM_CALLCONV XMColorSRGBToRGB( FXMVECTOR srgb )
 inline bool XMVerifyCPUSupport()
 {
 #if defined(_XM_SSE_INTRINSICS_) && !defined(_XM_NO_INTRINSICS_)
-#if defined(_XM_F16C_INTRINSICS_) || defined(_XM_AVX_INTRINSICS_)
-   int avxCPUInfo[4] = {-1};
-   __cpuid( avxCPUInfo, 0 );
+    int CPUInfo[4] = { -1 };
+    __cpuid(CPUInfo, 0);
+    if (CPUInfo[0] < 1)
+        return false;
 
-   if ( avxCPUInfo[0] < 1  )
-       return false;
-
-    __cpuid(avxCPUInfo, 1 );
+    __cpuid(CPUInfo, 1);
 
 #ifdef _XM_F16C_INTRINSICS_
-    if ( (avxCPUInfo[2] & 0x38000000 ) != 0x38000000 )
-        return false; // No F16C/AVX/OSXSAVE support
-#else
-    if ( (avxCPUInfo[2] & 0x18000000 ) != 0x18000000 )
-        return false; // No AVX/OSXSAVE support
+    if ((CPUInfo[2] & 0x38080001) != 0x38080001)
+        return false; // No F16C/AVX/OSXSAVE/SSE4.1/SSE3 support
+#elif defined(_XM_AVX_INTRINSICS_)
+    if ((CPUInfo[2] & 0x18080001) != 0x18080001)
+        return false; // No AVX/OSXSAVE/SSE4.1/SSE3 support
+#elif defined(_XM_SSE4_INTRINSICS_)
+    if ((CPUInfo[2] & 0x80001) != 0x80001)
+        return false; // No SSE3/SSE4.1 support
+#elif defined(_XM_SSE3_INTRINSICS_)
+    if (!(CPUInfo[2] & 0x1))
+        return false; // No SSE3 support  
 #endif
-#endif
-#ifdef _XM_SSE4_INTRINSICS_
-   int CPUInfo[4] = {-1};
-   __cpuid( CPUInfo, 0 );
 
-   if ( CPUInfo[0] < 1  )
-       return false;
+    // The x64 processor model requires SSE2 support, but no harm in checking
+    if ((CPUInfo[3] & 0x6000000) != 0x6000000)
+        return false; // No SSE2/SSE support
 
-    __cpuid(CPUInfo, 1 );
-
-    if ( (CPUInfo[2] & 0x80001) != 0x80001 )
-        return false; // Missing SSE3 or SSE 4.1 support
-#endif
-#if defined(_M_X64)
-    // The X64 processor model requires SSE2 support
     return true;
-#elif defined(PF_XMMI_INSTRUCTIONS_AVAILABLE)
-    // Note that on Windows 2000 or older, SSE2 detection is not supported so this will always fail
-    // Detecting SSE2 on older versions of Windows would require using cpuid directly
-    return ( IsProcessorFeaturePresent( PF_XMMI_INSTRUCTIONS_AVAILABLE ) != 0 && IsProcessorFeaturePresent( PF_XMMI64_INSTRUCTIONS_AVAILABLE ) != 0 );
-#else
-    // If windows.h is not included, we return false (likely a false negative)
-    return false;
-#endif
 #elif defined(_XM_ARM_NEON_INTRINSICS_) && !defined(_XM_NO_INTRINSICS_)
-#ifdef PF_ARM_NEON_INSTRUCTIONS_AVAILABLE
-    return ( IsProcessorFeaturePresent( PF_ARM_NEON_INSTRUCTIONS_AVAILABLE ) != 0 );
-#else
-    // If windows.h is not included, we return false (likely a false negative)
-    return false;
-#endif
+    // ARM-NEON support is required for the Windows on ARM platform
+    return true;
 #else
+    // No intrinsics path always supported
     return true;
 #endif
 }